EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix w8a8_int8 model shared experts fusion load weights error (#5120)

Browse Source
This commit is contained in:
lambert0312
2025-04-12 14:33:51 +08:00
committed by GitHub
parent 3c9740d200
commit 1b1b47a949
1 changed files with 18 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
python/sglang/srt/models/deepseek_v2.py
View File

@@ -1480,14 +1480,24 @@ class DeepseekV2ForCausalLM(nn.Module):
if self.n_share_experts_fusion is not None and self.n_share_experts_fusion > 0:
weights_list = list(weights)
weights_dict = dict(weights_list)
suffix_list = [
"down_proj.weight",
"down_proj.weight_scale_inv",
"gate_proj.weight",
"gate_proj.weight_scale_inv",
"up_proj.weight",
"up_proj.weight_scale_inv",
]
if self.quant_config.get_name() == "w8a8_int8":
suffix_list = [
"down_proj.weight",
"down_proj.weight_scale",
"gate_proj.weight",
"gate_proj.weight_scale",
"up_proj.weight",
"up_proj.weight_scale",
]
else:
suffix_list = [
"down_proj.weight",
"down_proj.weight_scale_inv",
"gate_proj.weight",
"gate_proj.weight_scale_inv",
"up_proj.weight",
"up_proj.weight_scale_inv",
]
names_to_remove = []
for moe_layer in tqdm(
range(